Memory Poisoning: The Shift to Persistent AI Threats

Prompt injection attacks are temporally bounded, but memory poisoning persists across sessions, allowing a single compromised memory to influence countless future decisions. As AI systems gain memory capabilities to improve usability—storing preferences, project context, and historical data—attackers gain persistent attack surfaces that single-session exploits never offered. This shift from stateless to stateful threats fundamentally changes the security model, moving threat boundaries from individual interactions to the entire timeline of the system's persistent state.